> There is no useful theory that allows us to predict the
> characteristics of the produced sequences from a set of possible
> seeds, so limiting the set of possible seeds is potentially dangerous.

I still find it difficult to understand where is the said danger. As you point out, a sequence of zeroes is a valid random sequence, and there is no reason to believe that a sequence of zeroes is more likely with a 256 bits seed, than with a 20 kbits seed (it might as well be less likely, for all we know).

We may as well bump it from 256 to 512 or 1024 bits, but 20 kbits sounds extremely unusual for a program to read from /dev/urandom at startup.